Systems and methods provide for generating personalized recommendations of digital content or application actions based on user context inferred from on device data. The system identifies a set of potentially relevant applications prior to receiving any user input using machine learning model that leverages current user context to assign likelihood scores to candidate applications. The most relevant applications are then surfaced to the user interface for immediate engagement. The system further collects one or more data streams from various applications, each stream representing user interactions. These streams are processed to extract contextual information in accordance with one or more predefined schemas that specify structured attributes. The resulting records are stored in a relational database which can be efficiently queried by authorized applications and services for personalized recommendations.